المفاهيم الأساسية
A novel image hashing framework, NeuroHash, that leverages hyperdimensional computing to enable spatial-aware and conditional image retrieval.
الملخص
The content presents a novel image hashing framework called NeuroHash that utilizes hyperdimensional computing (HDC) to enable spatial-aware and conditional image retrieval.
Key highlights:
- NeuroHash combines pre-trained vision models with HDC operations to encode spatial information into high-dimensional vectors, reshaping image representation.
- The framework allows dynamic hash manipulation for conditional image retrieval by controlling the weights on global and local features.
- NeuroHash outperforms state-of-the-art hashing methods on standard image retrieval benchmarks, demonstrating enhanced retrieval accuracy.
- The authors introduce a new evaluation metric, mAP@Kr, to measure the effectiveness of spatial-aware conditional image retrieval.
- Experiments on CIFAR-10 and MS COCO datasets validate the efficacy of the proposed approach.
The authors argue that NeuroHash breaks from traditional gradient-based training, offering a flexible and conditional image retrieval solution by seamlessly combining DNN-based neural and HDC-based symbolic models.
الإحصائيات
The MS COCO dataset contains 122,218 images from 80 categories, with a random sample of 5,000 images used as the query dataset and the remaining images as the retrieval set.
The CIFAR-10 dataset has 60,000 images distributed across 10 categories, with 100 images randomly chosen from each class to form the query dataset.
اقتباسات
"To resolve the above limitations of previous methods, we propose an innovative image hashing method employing Hyperdimensional Computing (HDC) [13] to facilitate image retrieval with spatial structural conditions that can be easily manipulated as illustrated in Figure 1."
"By combining DNN-based neural models with HDC-based symbolic models, our framework is capable of flexible hash value manipulation to have conditional image retrieval in a neuro-symbolic manner such as focusing on spatial information of a specific object."